You are on page 1of 52

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/226940429

Categoriality and Object Extraction in Cantonese Serial


Verb Constructions

Article  in  Natural Language & Linguistic Theory · August 2006


DOI: 10.1007/s11049-006-0005-3

CITATIONS READS

15 264

2 authors:

Elaine J. Francis Stephen Matthews


Purdue University The University of Hong Kong
35 PUBLICATIONS   278 CITATIONS    55 PUBLICATIONS   1,348 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Defaults and Prototypes View project

Lexical Categories View project

All content following this page was uploaded by Elaine J. Francis on 14 December 2017.

The user has requested enhancement of the downloaded file.


Natural Language & Linguistic Theory (2006) 24:751–801 © Springer 2006
DOI 10.1007/s11049-006-0005-3

ELAINE J. FRANCIS and STEPHEN MATTHEWS

CATEGORIALITY AND OBJECT EXTRACTION


IN CANTONESE SERIAL VERB CONSTRUCTIONS

ABSTRACT. The Cantonese ‘coverb’ construction, a serial verb construction in


which the first verb (the ‘coverb’) has a preposition-like meaning and function,
presents a challenge for theories of wh-dependencies and island constraints. Co-
verbs resist extraction of their objects by topicalization or relativization, a fact
which has often been explained in terms of a preposition-stranding constraint in
accounts of similar facts in Mandarin. However, Cantonese coverbs display the
morphosyntactic properties of verbs, suggesting that they cannot be prepositions.
In this paper, we propose that coverbs are verbs, and that the relevant extrac-
tion constraint is a kind of adjunct island constraint. This proposal is supported
with experimental evidence from a sentence judgment task. Two key findings are
as follows: (1) listeners judged extraction from a coverb phrase as significantly
less acceptable than extraction from a simple clause; (2) listeners judged sen-
tences both with and without aspectual marking (verbal morphosyntax) on the
coverb as highly acceptable. Together, these findings support our proposal that
coverbs are verbs (not prepositions) and that coverb phrases form a kind of
adjunct island. However, we show that existing adjunct island conditions (such as
the CED) are not adequate to account for our data. Following Hawkins’ (1999)
processing-based theory of filler-gap dependencies, we propose a simple, language-
specific formulation of the extraction constraint, and we argue that this constraint
is more generally motivated by a processing principle called Avoid Competing
Subcategorizers – one of the same principles that motivates preposition-stranding
constraints in other languages. Thus, although object extraction is prohibited by
a kind of adjunct island constraint, the function of the constraint in processing
efficiency is similar to that of a preposition-stranding constraint.

 This work has been substantially supported by two grants from the Univer-
sity Research Committee of the University of Hong Kong. An earlier version of
this paper was presented at the Linguistic Society of America annual meeting in
San Francisco, 2002. We thank Alex Francis for assistance with experimental design
and statistics, and Kawaii Yeung, Michelle Li and Helen Ching for helping collect
the experimental and corpus data. For judgments on Cantonese and Mandarin we
are grateful to Virginia Yip, Richard Wong, Nicole Li, Tommy Leung and Liang
Yuan. We thank Paul Law for discussion of theoretical issues, and Peter Culicover,
Jerry Sadock and two anonymous NLLT reviewers for insightful comments on ear-
lier versions. Any remaining errors are our own.
752 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

1. I NTRODUCTION : THE P ROBLEM OF C ATEGORIALITY


AND O BJECT E XTRACTION

The term ‘serial verb construction’ (SVC) is used to refer to con-


structions containing multiple verbs within a single clause, as in the
following example from Yoruba:
(1) Bólá ra bàtà fún Fémi.
Bola bought shoes give Femi.
‘Bola bought shoes for Femi.’ (Lawal 1993: 186)
SVCs pose a number of interesting problems for theories of syn-
tax and semantics, in particular with respect to categoriality (Ross
1991, McCawley 1992; Hagemeijer 2001), word order and constituency
(Dai 1990; Déchaine 1993a,b; Carstens 2002), movement (Lawal 1993,
Law 1996, Veenstra 2000), argument structure (Finney 1992, Collins
1997) and predicate composition (Andrews and Manning 1999).
Lord (1993) observes that a particular subset of SVCs has proven
especially troublesome for linguistic theory and description: “These
are the constructions in which one of the verbs is defective in some
respect, such as. . . failure to take the usual verb inflections or nega-
tion affixes, or showing unexpected syntactic properties, for example
with respect to movement” (Lord 1993: 3). In this paper, we explore
a set of puzzling facts regarding categoriality and object extraction
in one such SVC in Cantonese: the ‘coverb’ construction.
In Chinese linguistics, the traditional term ‘coverb’ refers to a set
of transitive verbs that typically occur in V1 position of an SVC of
the form V1 NP1 V2 (NP2). A few of the most common Cantonese
coverbs are given in (2):
(2) a. tung4 ‘with’ g. wai4 ‘for the
sake of’
b. hai2 ‘(be) at’ h. jau4 ‘from’
c. gan1 ‘follow, with’ i. deoi3 ‘toward’
d. doi6 ‘replace, in place of’ j. ging1 ‘pass, via’
e. bong1 ‘help, for’ k. hoeng3 ‘toward,
facing’
f. wan2 ‘use, seek’ l. ziu3 ‘following’
Coverbs have relational meanings indicating location, direction, instru-
ment, beneficiary, and similar notions, and they are often translated as
prepositions in English. The phrase containing the coverb precedes and
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 753

modifies the main VP. In (3), the coverb phrase bong go go jan ‘for that
person’ precedes and modifies the main VP zou-je ‘work’.1
(3) ngo bong go go jan zou-je .
I help that CL person do-things
‘I work for that person.’
A peculiar property of coverbs is that they resist extraction of their
objects by relativization or topicalization, as shown for relativiza-
tion in (4).
(4) ?*[[ngo bong i zou-jeRC ] go go janiNP ]
I help do-things that CL person
Intended: ‘the person that I work for ’
In contrast, the main verb (V2) of a coverb construction allows
extraction, as shown by the relative clause in (5):
(5) [[ngo bong keoi zou i RC ] go di jei NP ]
I help 3sg do that CLPL thing
‘The things I do for him’
In studies on Mandarin Chinese, which has a similar coverb construc-
tion, the extraction facts are commonly attributed to a constraint on
‘preposition-stranding’, whereby the objects of prepositions cannot
be moved from the position directly following the preposition (Huang
1982, McCawley 1992, Li 1990, Zhang 1990). According to this view,
coverbs are prepositions when occurring in the coverb construction,
although some of them may be used as verbs in other contexts. Such a
view is particularity problematic in Cantonese, however, because co-
verbs display verbal properties even when they are used like preposi-
tions. For example, Cantonese coverbs can occur with aspect marking
(6) and various verbal particles (7–8).
(6) ngo tung-gwo keoidei king-gai (aspect marking)
I with-ASP them chat
‘I’ve chatted with them.’

1 Cantonese examples are given in the JyutPing romanization developed by the


Linguistic Society of Hong Kong (Kwok et al. 1997). Tone marks are omitted as
they are not relevant to the discussion. The abbreviations used in our Cantonese
examples are as follows: ASP aspect, CL classifier, PRT particle, MOD modal and
REL relative marker.
754 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

(7) ngo m tung dak keoidei king-gai (modal DAK)


I not with MOD them chat
‘I can’t chat with them.’
(8) ngo tung saai keoidei king-gai (SAAI quantification)
I with all them chat
‘I chat with all of them.’
How can the verbal properties of coverbs be reconciled with the
preposition-like constraint on extraction? In Section 3, we argue that
Cantonese coverbs are verbs with somewhat non-prototypical syn-
tactic and semantic functions, not prepositions. We propose that the
relevant constraint on extraction is an adjunct island condition along
the lines of Huang’s Condition on Extraction Domain (CED), which
disallows extraction from phrases or clauses functioning as adjuncts.
However, the data present some apparent problems for this approach.
One problem is that coverb phrases are not straightforwardly classi-
fied as adjuncts: coverbs allow aspect marking and verbal particles
(see (6–8) above), which are often assumed to be indicators of head-
edness. A further problem is that extraction seems less acceptable
with some coverbs than with others, even though all coverbs appear
to occur in the same syntactic configuration.
To shed light on these problems, we conducted a sentence-judgment
experiment in which six coverbs were tested along three parameters:
(1) 2-verb sentence (coverb construction) vs. 1-verb sentence (basic
clause); (2) aspect vs. no aspect; (3) extraction vs. no extraction. The
results of the experiment, discussed in Section 4, support our pro-
posal that coverbs are verbs, and that the constraint on extraction is
an adjunct island condition of some kind. Furthermore, the results
suggest that aspect marking, while relevant for indicating the cate-
gory status of coverbs, is irrelevant for determining which verb is the
head of the SVC. Finally, the results show that object extraction is
judged to be less acceptable with some coverbs than with others: we
suggest that this variability may be due to usage-based factors which
interact with the adjunct island constraint.
Having established the plausibility of an adjunct island constraint
in Sections 2–4, we address the theoretical problem of how best to
formulate the relevant constraint. In Section 5, we show that the
CED and other existing formulations of the adjunct island constraint
are fraught with technical difficulties when applied to the coverb
construction. We also observe that island constraints are subject
to a great degree of cross-linguistic and intra-language variability,
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 755

suggesting that any formulation of the adjunct island constraint is


unlikely to hold true for all languages, or even for all types of adjuncts
within a single language. Following the approach of Hawkins (1999),
we argue that the relevant constraint is best defined in language-spe-
cific, construction-specific terms, but at the same time motivated by a
general processing principle, Avoid Competing Subcategorizers. The
same processing principle also motivates preposition-stranding con-
straints in other languages, thus capturing the insight that the Can-
tonese adjunct island constraint is similar to a preposition-stranding
constraint, while acknowledging that it cannot be formulated as such
in relation to items that are not prepositions.
Before presenting our proposal for Cantonese, we will first briefly
discuss the better-studied coverb construction in Mandarin, and
how this construction has been analyzed in previous work.

2. P REVIOUS S TUDIES OF M ANDARIN

The Mandarin coverb construction is similar to the Cantonese co-


verb construction in most respects. In example (9) from Mandarin,
the first verb gei ‘give, for’ is the coverb and the second verb zuo
‘make’ is the main verb. The coverb phrase gei wo ‘for me’ precedes
and modifies the main VP zuo jiaozi ‘make dumplings’.
(9) mama gei wo zuo jiaozi
mother for I make dumpling
‘Mother made dumplings for me.’
(Li and Thompson 1981: 358)
As in Cantonese, all coverbs display a subset of verbal properties,
including the ability to function as verbs in yes-no (A-not-A) ques-
tion formation, as in (10). However, most of them are grammatically
defective, for example with respect to aspect marking, as in (11).
(10) mama gei bu gei wo zuo jiaozi?
mother for not for I make dumpling
‘Did Mother make dumplings for me?’
(11) mama gei (*le) wo zuo jiaozi
mother for (ASP) I make dumpling
‘Mother made dumplings for me.’
The defective properties of coverbs, along with their preposition-
like meanings and functions, lead to the appearance of a mixed
756 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

category status. Hence, Mandarin coverbs are often described as hav-


ing grammatical properties of both verbs and prepositions. Li and
Thompson (1981: 360) also observe that coverbs vary in their dis-
tributional behavior, with some behaving more like verbs and others
more like prepositions. For example, some coverbs can, like ordinary
verbs, occur independently as the main predicate of a clause (e.g.,
zai ‘at’), while others can occur only with other verbs in a coverb
construction (e.g., cong ‘from’).
From a diachronic perspective, it appears that the defective
properties of coverbs are the result of grammaticalization along a
well-attested cline from main verb to serial verb to preposition, as
described in Lord (1993), while the grammatical variation among
different coverbs is due to some items having progressed further in
this historical transition than others (Li and Thompson 1981: 360).
What does this mean for a synchronic analysis of their category
status? The simplest solution would be to assume that coverbs are
simply verbs, and that their defective properties are specified in their
individual lexical entries. This solution accommodates the verbal
properties of coverbs and allows for variability in their grammati-
cal behavior. However, the object extraction facts seem less amenable
to this sort of analysis, since constraints on extraction are normally
considered to result from general principles of grammar rather than
from the lexical properties of certain verbs. In previous studies, these
facts have been handled in one of two ways: (1) a constraint on
“preposition stranding”, and (2) a constraint on adjunct islands (or
similar structural constraint). According to the first approach, the
inability of Mandarin coverbs to be ‘stranded’ under movement con-
firms their status as prepositions, despite whatever properties they
share with verbs (Huang 1982, Li 1990, Zhang 1990, McCawley
1992). This approach derives its generality from a well-known univer-
sal tendency. In most languages that have prepositions, a preposition
cannot be left ‘stranded’ when its object is extracted, as in (12) from
French.2
2 We note that P-stranding occurs productively in a small minority of the world’s
languages, including English, Norwegian, Danish, and the West African languages
Vata and Gbadi (Koopman 1984). One widely-discussed proposal is that the prepo-
sition undergoes reanalysis to become part of a complex verb, which is then able to
assign objective case (Hornstein and Weinberg 1981: 60), as in (i):
(i) the girl he [VP [V spoke of] ]
Levine (1984), Baltin and Postal (1996), and others have shown empirical and
theoretical problems with a reanalysis account of P-stranding, and various other
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 757

(12) *la fille qu’il parlait de


the girl who he spoke of (Hornstein and
Weinberg 1981: 57)
This generalization is captured in various ways by different authors.
Most importantly for our purposes, constraints on preposition strand-
ing involve a grammatical distinction of some kind between transi-
tive verbs and prepositions. For example, Hornstein and Weinberg
(1981) propose a general constraint disallowing oblique case-marking
of traces: whereas transitive verbs assign objective case to their objects,
and thus allow extraction, prepositions assign oblique case to their
objects, and thus disallow extraction (‘P-stranding’).3
In analyses which assume a preposition-stranding account of
extraction, the verbal properties of some coverbs are handled in the
lexicon, either by listing some coverbs exclusively as verbs, or by
providing a dual listing of some coverbs as both verbs and preposi-
tions. McCawley (1992: 222) observed that most Mandarin coverbs
disallow extraction of their objects, but some marginally allow it.
He takes such data as evidence that some items traditionally classi-
fied as coverbs are prepositions (i.e., those that disallow stranding),
while others are verbs. Zhang (1990: 313) proposes a similar analysis
involving a dual lexical listing for those coverbs that can also occur
as main verbs. According to this approach, when a coverb occurs
in a coverb construction modifying the main VP, it is analyzed as
a preposition, and extraction of the object is ruled out accordingly;
when it occurs as the sole predicate of its clause, it is analyzed as a
verb. Although the analysis of (certain) Mandarin coverbs as prep-
ositions provides a general account of the constraint on extraction,
and is widely accepted in generative studies of Chinese syntax, it
ignores the grammatical properties that coverbs share with ordinary
verbs even in their “prepositional” usages, such as A-not-A ques-
tion formation (possible with all coverbs) and aspect marking (pos-
sible with some coverbs). This problem led Ross (1991), Law (1996),
and Li (1993) to propose that the relevant constraint is independent

Footnote 2 continued
accounts have been proposed. For example, Jones (1987) and Newmeyer (1998a)
argue that P-stranding is allowed only in languages for which prepositions are
proper governors. Further discussion of this issue is beyond the scope of the current
paper.
3 Other authors, such as Baltin (1978), have described the P-stranding constraint
in terms of Subjacency, proposing that PP and S are bounding nodes, but VP is not.
758 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

of the category status of coverbs. Thus Law (1996) proposes that


extraction is ruled out because coverb phrases function as adjuncts
to the main VP. For example, in (13), the verb qie ‘cut’ is the head
verb, and the coverb phrase na dao ‘hold knife’ is an adjunct VP.
(13) ta na dao qie-le rou
he hold knife cut-ASP meat
‘He cut the meat with a knife’
Law (1996: 218–220) argues that only the head verb allows extraction
of its object by topicalization or relativization, while the non-head
verb in the adjunct VP does not, as illustrated with respect to rela-
tivization in (14a–b).
(14)a. [[ta na dao qie-le iRC ] de roui NP ]
he hold knife cut-ASP REL meat
‘the meat that he cut with a knife’
b. *[[ta na i qie-le rouRC ] de daoi NP ]
he hold cut-ASP meat REL knife
Intended: ‘the knife that he cut the meat with’
He then attributes the ungrammaticality of extraction in (14b) to
Huang’s CED (Huang 1998: 359).
(15) Condition on Extraction Domain
A phrase A may be extracted out of a domain B only if
B is properly governed.
Although the definition of proper government is formulated differ-
ently by different authors, the basic idea is that phrases and clauses
functioning as complements are properly governed by the lexical head
that subcategorizes for them, whereas phrases and clauses function-
ing as adjuncts or subjects are not properly governed by any lexical
head. The CED is usually invoked to account for the ungrammatical-
ity of sentences in which a NP is extracted from a subordinate clause
functioning as an adjunct, as in the following English example:
(16) *Whoi did Mary cry after John hit i ?
(Huang 1998: 358)
Law reasons that if the coverb phrase functions as an adjunct to the
main VP, it is also an island to extraction. Consistent with this analysis,
Law (1996: 219) also shows that extraction from VP1 is allowed in other
kinds of SVCs for which V1 is clearly the main verb, as in (17):
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 759

(17) ta song-le lai de xiangzi


he send-ASP come REL suitcase
‘the suitcase that he sent over’ (Law 1996: 218)
Déchaine (1993b) provides a similar analysis of Mandarin coverb
phrases as adjuncts (although she does not discuss extraction),
and provides tree diagrams to illustrate her analysis. Following
Déchaine, we may represent the Mandarin coverb construction and
the position from which extraction is disallowed as in (18) below.
Here, we use subscripts and boldface type to indicate that, in the
case of the sentence in (13) above, V2 is the head of the entire SVC
(the matrix VP), while V1 is the head of an adjunct VP.4
(18)
S

NP VP 2

ta
he
VP1 VP 2

V1 NP V2 NP

na dao qie-le rou


hold knife cut-ASP meat

From the diagram in (18), it appears that there is in fact a


CED violation when extracting the object of na ‘hold’, since there
is no lexical item governing the VP na dao ‘hold knife’. However,
extracting the object of qie ‘cut’, which is grammatical as shown
in (14a) above, would also appear to violate the CED, since there
is apparently no lexical item governing the matrix VP either. As
Chomsky (1986) recognizes, the CED is in general problematic: the
matrix VP is never properly governed by a lexical head, and yet

4 Note that although we analyze the matrix clause as belonging to the category
S, analyzing it instead as a functional projection (e.g., IP) would not affect the anal-
ysis of the extraction facts.
760 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

extraction of the matrix VP’s object is, in general, perfectly gram-


matical. Despite this technical difficulty, some kind of adjunct island
constraint remains a plausible hypothesis. We will return to the
problem of formulating this constraint in Section 5.
In sum, this section has discussed two general approaches to
the problem of categoriality and object extraction that have been
applied in previous work on Mandarin coverb constructions. The
first approach says that when coverbs are used in the coverb con-
struction, they are prepositions. Their incompatibility with extrac-
tion arises from a general constraint on preposition stranding,
which is attested in most languages that have prepositions. How-
ever, this approach ignores the properties that coverbs share with
verbs even in their “prepositional” usages. The second approach for-
mulates the constraint on extraction as an adjunct island constraint
which applies regardless of the syntactic category status of coverbs.
This approach leaves open the possibility that some or all coverbs
are verbs, thus accommodating their verbal properties. In the follow-
ing section, we argue that our Cantonese data are best understood
using a version of the second approach.

3. C ATEGORIALITY AND O BJECT E XTRACTION IN C ANTONESE

In this section, we elaborate on the problem of categoriality and


object extraction in Cantonese and propose a theoretical account of
this problem. The fact that coverb phrases disfavor extraction of their
objects is often used as evidence either that coverbs are prepositions,
or that coverb phrases are adjuncts to the main VP. To determine the
best analysis of the extraction constraint, however, we need evidence
independent of extraction. In Sections 3.1–3.2, we present evidence
that coverbs are verbs, not prepositions, and that coverb phrases
are adjuncts. Such evidence argues against a preposition-stranding
account of the extraction facts and in favor of an adjunct island
account. Finally, in Section 3.3, we discuss some remaining problems
for the analysis. One problem is that the adjunct status of coverb
phrases is not completely straightforward: aspect marking and other
morphosyntactic features, which are often assumed to indicate head-
edness, can occur on either the coverb or the main verb. In addi-
tion, there seems to be an interaction between aspect marking and
extraction, as well as variation among individual coverbs regarding
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 761

the acceptability of object extraction. These problems motivate the


experimental study discussed in Section 4.

3.1. Category Status of Cantonese Coverbs


Cantonese coverbs share most of the syntactic and morphosyntactic
properties of ordinary verbs. Similar to Mandarin coverbs, Canton-
ese coverbs act as verbs with respect to A-not-A question formation
and negation with m-, as shown in (19a–b):
(19)a. lei tung-m-tung ngodei sik faan aa?
you with-not-with us eat rice PRT
‘Will you be eating with us?’
b. ngo m tung leidei sik faan
I not with you.PL eat rice
‘I won’t be eating with you’
Also like Mandarin coverbs, most Cantonese coverbs can function
as the main predicate of a simple clause. The few that generally can-
not include tung4 ‘with’, and jau4 ‘from’:5
(20) *ngo jigaa tung go di jan
I now with those CL people
Intended: ‘I’m with those people now.’
(21) *ngo jau nidou
I from here
Intended: ‘I (come) from here.’

5 In fact tung4 may occur as sole predicate in at least two kinds of circum-
stances: with progressive aspect marking (i) and when the implied action corre-
sponding to the missing main verb is clear from the context (ii):

(i) Gaaze tung-gan go go laamzai wo.


sister with-ASP that CL boy PRT
‘Sister is hanging out with that boy.’

(ii) Ngo tung lei laa!


I with you PRT
‘Let me (go) with you.’ (e.g. when getting into a taxi)
762 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

Most coverbs, however, can readily function as main predicates, for


example hai2 ‘(be) at’ and doi6 ‘replace’:

(22) keoi zung hai ukkei


S/he still at home
‘She’s still at home.’
(23) ngo doi ni go jan
I replace this CL person
‘I’m (acting) in this person’s place.’

Finally, Cantonese coverbs display the morphosyntactic trappings


of verbs to a greater extent than Mandarin coverbs. Although
aspect is typically marked on the main verb (V2) of a coverb con-
struction, coverbs in V1 position freely allow aspect marking as
well as a range of post-verbal particles indicating notions such
as modality and quantification (see examples 6–8 in Section 1
above). Thus, the available evidence indicates that coverbs are a
non-homogeneous set of grammaticalized verbs. Their verbal prop-
erties follow from their categorization as verbs, and the defective
properties of some coverbs may plausibly be listed in their lexical
entries.
As in Mandarin, Cantonese coverbs disallow extraction of their
objects when used in the coverb construction. However, this does
not necessitate an analysis of coverbs as prepositions. As we argue
in the following sections, the preposition-like properties of co-
verbs follow from their non-prototypical meanings and functions
within the coverb construction, while the constraint on extrac-
tion is best understood as a constraint on extraction from adjunct
islands.

3.2. Constituency and Asymmetrical Structure


of the Coverb Construction
A proper formulation of the extraction facts first requires an anal-
ysis of the constituent structure, semantic structure, and headedness
of the coverb construction. As a working hypothesis, let us suppose
that the Cantonese coverb construction has the constituent structure
proposed by Déchaine (1993b) for Mandarin, as given in (18) above.
A Cantonese example is given in (24):
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 763

(24)
S

NP VP 2

keoi
s/he
VP 1 VP 2

V1 NP V2 NP

tung go go laamzai haang gungsi


with that CL boy walk shop

This diagram indicates that the coverb and its object form a constit-
uent, that the main verb and its object form a separate constituent,
and that the entire SVC forms a constituent consisting of the coverb
phrase and the main verb phrase. This analysis is confirmed by the
coordination test, as shown in examples (25a–c).
(25) a. keoi [tung go go laamzai] tungmaai
s/he with that CL boy and
[tung go go leoizai] haang gungsi.
with that CL girl walk shop
‘She goes shopping with that boy and with that girl.’
b. keoi tung go go laamzai
s/he with that CL boy
[haang gungsi] tungmaai [gong je].
walk shop and talk things
‘She goes shopping and discusses things with that boy.’
c. keoi [tung go go laamzai haang gungsi]
s/he with that CL boy walk shop
tungmaai [tung go go leoizai gong je].
and with that CL girl talk things
‘She goes shopping with that boy and discusses things
with that girl.’
Binding facts lend further support for the constituent structure in
(24). Assuming the standard c-command condition on binding, we
764 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

should expect that only the subject is a possible antecedent to a


reflexive pronoun following the main verb. This is in fact what we
find: when the main verb takes a reflexive pronoun as its object,
the object of the coverb cannot serve as its antecedent. In (26), for
example, only the subject can be the antecedent of zigei ‘self’:6,7
(26) keoii tung go go laamzaik
s/he with that CL boy
gong zigeii/∗k ge mantai
talk self POSS problem
‘She talks to that boy about her own problems.’
NOT: ‘She talks to that boy about his problems.’
The diagram in (24) also shows an asymmetrical structure for
the coverb construction. As indicated by the boldface type and
numbering, V2 is the head of the entire SVC, while VP1 is an
adjunct to VP2. What evidence is there for an asymmetrical struc-
ture? First, the coverb construction is clearly asymmetrical with
respect to semantics. The coverb phrase has the semantic function
of a modifier to the main verb phrase, which is the main predi-
cate of the clause. This is consistent with the observation that the
coverb phrase is always optional, whereas the main verb phrase is
required to maintain the basic meaning of the sentence. A coverb
phrase such as hai go gaan sezilau ‘in that office’ in (27a) can be
omitted, as in (27b), without changing the basic propositional con-
tent of the sentence. The coverb phrase in (27a) simply contains
additional information about where the meeting is being held.
(27)a. A-Maa jigaa (hai go gaan sezilau) gin haak
Mom now (at that CL office) see client
‘Mom’s meeting clients (in that office) right now.’

6 Note that zigei ‘self’ is a gender-neutral reflexive pronoun which can potentially
refer to either a male or a female. Peter Culicover (p.c.) notes that a semantic theory
of binding as proposed in Culicover and Jackendoff (2005) would make the same
predictions, assuming that the tung-phrase is an adjunct in Conceptual Structure.
7 We note that example (26) is inconsistent with a structure such as that
proposed by Rhys (2000: 79) for the Mandarin coverb construction, whereby
the coverb’s object is analyzed as a subject occurring in Spec, VP position and
c-commanding the main verb. We leave open the question of whether Rhys’ pro-
posal works for Mandarin, since the binding facts are somewhat different in this
language.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 765

b. A-Maa jigaa gin haak


Mom now see client
‘Mom’s meeting clients right now.’

c. A-Maa jigaa hai go gaan sezilau


Mom now at that CL office
‘Mom’s in that office right now.’
Although the main verb phrase gin haak ‘see client’ can also be
omitted, as in (27c), the sentence now expresses a completely differ-
ent proposition: whereas (27b) expresses the proposition that Mom
is meeting clients, (27c) expresses the proposition that Mom is in the
office.
Secondly, it is clear that the coverb construction is not a coordi-
nate structure. If it were, we would expect the Coordinate Structure
Constraint (Ross 1967) to rule out extraction from either one of the
conjuncts. But as shown in Section 1, example (5), extraction from
VP2 is perfectly acceptable, while extraction from VP1 is not. This
is again consistent with an analysis of VP1 as an optional adjunct
to VP2.
Thirdly, the coverb phrase must precede the main verb phrase.
This follows the general pattern found in Cantonese, in which
VP-adjuncts precede rather than follow the head verb. Coverb
phrases such as gan go go sinsaang ‘follow(ing) that teacher’ must
precede the main verb (28a–b), just like VP-modifying adverbs such
as ji5cin4 ‘formerly’ (29a–b).8
(28) a. A-Man gan go go sinsaang hok kam
Ah Man follow that CL teacher learn piano
‘Ah Man is learning the piano with that teacher’
b. *A-Man hok kam gan go go sinsaang
Ah Man learn piano follow that CL teacher
(29) a. A-Man jicin hok kam
Ah Man formerly learn piano
‘Ah Man was learning the piano before.’
b. *A-Man hok kam jicin
Ah Man learn piano formerly

8 (28b) has a grammatical reading in which [hok kam] ‘learn piano’ is a sec-
ondary topic, i.e. ‘as far as learning the piano is concerned, she studies with that
teacher’. On this construction see Matthews and Yip (1994: 75).
766 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

Thus, both syntactically and semantically, the coverb phrase is simi-


lar in function to an optionally expressed VP-modifying adverb. All
of this suggests that the coverb construction has an asymmetrical
structure in which VP2 is the main predicate of the clause, and VP1
is an adjunct to VP2 (see also Matthews 2006).
Finally, let us return to the question of category status. Although
Cantonese coverbs share most of the syntactic and inflectional
properties of transitive verbs, coverbs (as used in the coverb con-
struction) are atypical of verbs with respect to their semantic
function as modifiers, and with respect to their syntactic function as
VP-adjuncts. These unusual properties are consistent with the general
observation that category members tend to vary in their grammat-
ical behavior, with some items behaving in a more “prototypical”
manner than others (Croft 1991, Hopper and Thompson 1984).9
Crosslinguistically, verbs are particularly prone to variation in their
argument structure and selection properties, as well as in their com-
patibility with various markers of tense, aspect, and mood. Further-
more, as we argue in Francis and Matthews (2005), the category
‘verb’ has an especially broad membership in Cantonese and sub-
sumes a number of items that might translate as adjectives or prep-
ositions in other languages. We conclude that coverbs are (partially
grammaticalized) verbs, some of which are defective in their distri-
butions, and all of which are commonly used as adjuncts to the
main verb within a SVC.

3.3. Aspect Marking, Extraction, and Headedness


We have argued in Sections 3.1–3.2 that coverbs are verbs which
can function as adjuncts to the matrix VP in a SVC. This evi-
dence supports the hypothesis that coverbs resist extraction of their
objects due to an adjunct island constraint (as advocated by Law
1996), rather than a preposition-stranding constraint. Thus, leaving
aside the exact formulation of the adjunct island constraint until

9 Hopper and Thompson (1984) and Croft (1991) show that this variation is sys-
tematic in nature, suggesting a prototype structure for lexical categories. But see
Baker (2003) and Newmeyer (1998b) for criticisms of prototype theories and for
proposals that variation among category members can be dealt with in terms of
independent principles. For the purposes of the present analysis, we remain neu-
tral as to whether categories literally have a prototype structure or not. See Francis
(2005) for discussion.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 767

Section 5, we can assume that there is a constraint barring the


extraction out of a phrase functioning as an adjunct to another
phrase. This constraint accounts for the marginal ungrammaticality
of extraction of the object of V1, as illustrated in (4) in Section 1.
There is, however, one kind of evidence which calls into ques-
tion the analysis of coverb phrases as adjuncts. If V2 is in fact the
head of the SVC, we might also expect that any morphosyntactic
features of the SVC would be manifested on V2 rather than (or in
addition to) V1 (Li 1991: 112). This is largely true for Mandarin: if
aspect is marked at all in a coverb construction, it normally must be
marked on V2, the main verb (Li 1991; Law 1996; Déchaine 1993b).
Following Chan’s (2000: 92) GPSG analysis of one type of SVC in
Mandarin (i.e., SVCs with a ‘circumstance’ interpretation), we might
suppose that the aspect feature (a type of head feature) is instan-
tiated both on the head verb of the SVC and on the VP dominat-
ing the entire SVC. In the following example, the aspect feature is
instantiated on V2 (taolun ‘discuss’) and on the higher VP, given
in brackets. This is sometimes described as percolation of head fea-
tures from the head verb to the higher VP (e.g., Zwicky 1985).
(30) women [VP kai hui taolun-guo nei ge wenti]
we open meeting discuss-ASP that CL problem
‘We had discussed that problem in the meeting.’
Such an analysis is consistent with the semantic effect of the aspect
marker, which takes scope over the entire sentence, not just VP2.
However, Chan (2000: 89) observes that in some cases, it is also pos-
sible to mark aspect (specifically, experiential aspect) on V1 without
any change in meaning, as in the following example:
(31) women kai-guo hui taolun nei ge wenti
we open-ASP meeting discuss that CL problem
‘We had discussed that problem in the meeting.’
Again, VP1 apparently modifies VP2, and experiential aspect takes
semantic scope over the entire sentence. Presumably for these rea-
sons, Chan still analyses V2 as the head of the SVC bearing the
aspectual head feature.
A similar problem arises with respect to the Cantonese coverb
construction. Although aspect is usually indicated on V2, as in
(32a), marking experiential aspect on V1 is also acceptable, as in
(32b).
768 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

(32)a. ngo tung go go jan touleon-gwo je.


I with that CL person discuss-ASP things
‘I’ve had discussions with that person.’ (aspect on V2)
b. ngo tung-gwo go go jan touleon je.
I with-ASP that CL person discuss things
‘I’ve had discussions with that person.’ (aspect on V1)
Regardless of which verb takes the aspect marking, the coverb
phrase still semantically modifies the main VP. There are two likely
possibilities here: (1) the position of aspect marking is not neces-
sarily indicative of headedness, and V2 is always the head in a co-
verb construction as assumed by Chan (2000) for (31); or (2) the
position of aspect marking is indicative of syntactic headedness, but
semantically, V2 remains the main predicate. This second option
is not discussed in the literature, but is plausible in the light of
recent work by Culicover and Jackendoff (1997, 2005) and Yuasa
and Sadock (2002) demonstrating that relationships such as subor-
dination and coordination can be independently expressed in syn-
tax and semantics. For example, Yuasa and Sadock (2002) present
persuasive evidence that in so-called te-coordination in Japanese, the
first conjunct is subordinate to the second conjunct in syntax, but
the two conjuncts are coordinate in semantics. A similar analysis of
the Cantonese coverb construction might say that an aspect-marked
coverb is the syntactic head, but still functions as a semantic mod-
ifier of the following VP.
The evidence adduced in Section 3.2 in favor of an adjunct anal-
ysis of VP1 would seem to favor the first option. That is, if the
position and the optionality of the coverb phrase are evidence in
favor of an adjunct analysis of VP1, then the fact that aspect can be
marked on either verb would appear to be independent of headed-
ness: VP1 behaves like an adjunct in other respects, regardless of the
position of aspect marking. We would predict, then, that if aspect is
irrelevant to headedness, and if extraction is ruled out by an adjunct
island constraint, then sentences involving extraction from VP1 of a
coverb construction should be (equally) bad whether or not aspect
is marked on V1 (the coverb). On the other hand, if aspect can only
be marked on the head verb, then extraction from VP1 should be
possible in sentences for which aspect is marked on V1, regardless
of the semantic function of VP1.
A preliminary survey showed a possible interaction between
aspect marking and extraction. According to judgments we collected
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 769

in an informal survey of three native Cantonese speakers, sen-


tences in which the object of a coverb was extracted as in (33a)
were judged somewhat more acceptable when an aspect marker was
attached to the coverb as in (33b):
(33)a. *[[ngo bong i zou jeRC ] go go janiNP ]
I help do things that CL person
hou sing ge
very smart
Intended: ‘The person I work for is very smart.’
b. ?[[ngo bong-gwo i zou jeRC ] go go
I help–ASP do things that CL
janiNP ] hou sing ge
person very smart
Intended: ‘The person I’ve worked for is very smart.’
These judgments suggest that aspect marking may improve the
acceptability of extraction from VP1 of a coverb construction. If
confirmed, such findings would suggest that aspect marking causes
listeners to interpret V1 as the syntactic head of the SVC, thus mak-
ing the adjunct island constraint inapplicable. This would then sup-
port an analysis whereby the adjunct island constraint is specified
only at the level of syntax and fails to apply to phrases that are
semantic adjuncts (modifiers) but syntactic heads.
Due to the subtlety and variability of the judgments obtained
on the informal survey, a more carefully controlled study was nec-
essary. A sentence judgment experiment was therefore designed to
test for possible interactions between aspect marking and extrac-
tion. The main questions that we set out to test were as follows:
(1) Do coverbs consistently allow aspect marking in the coverb con-
struction? If so, this would support our proposal that coverbs are
verbs, not prepositions. (2) Does extraction of the coverb’s object
consistently result in lowered acceptability for sentences involving
the coverb construction? If so, this would support our proposal that
an adjunct island constraint bars extraction from a coverb phrase
functioning as an adjunct. (3) Does aspect marking on the coverb
improve the acceptability of sentences for which the coverb’s object
is extracted? If so, this would support our hypothesis that aspect
marking on the coverb may cause it to be interpreted as the head of
the SVC, making the adjunct island constraint inapplicable in such
770 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

cases. Specific hypotheses and predictions based on these questions


are detailed below in Section 4.2.

4. S ENTENCE J UDGMENT E XPERIMENT

This experiment investigated the effects of object extraction from


VP1 on the acceptability of sentences involving the coverb con-
struction. In the experiment, 40 participants heard 56 sentences and
judged each of them on a scale from 1 (‘good, completely accept-
able’) to 4 (‘bad, completely unacceptable’). Three linguistic fac-
tors were manipulated: (1) presence vs. absence of aspect marking
on V1; (2) choice of verb in V1 position; and (3) sentence type.
A factorial design was used. Aspect was a between-groups factor
with two levels (aspect/no aspect), while verb and sentence type were
within-group factors with six levels (six different verbs) and four lev-
els (four different sentence types), respectively. The sentence types
included simple one-verb clauses with and without extraction of
their objects (control sentences), and two-verb clauses (coverb con-
structions) with and without extraction of the object of V1 (test sen-
tences). The same six verbs were used in the simple sentences and in
the V1 position of the two-verb sentences. Details of the experimen-
tal design are given in Section 4.3 (Methods) below.

4.1. Motivation for Experimental Methodology


Much of the work in generative linguistics conducted during the last
40 years relies on native-speaker intuitions about the acceptability
of sentences with particular grammatical characteristics. The use of
such sentence judgments has made it possible to obtain evidence for
what is and is not grammatical in a language – a key goal of gen-
erative linguistic research. While other data sources, such as speech
or text corpora, can also be very useful, corpus examples give no
direct evidence for the ungrammaticality of a particular structure.
In addition, the use of sentence judgments allows researchers to test
pairs or sets of sentences that differ only along the dimension under
investigation, thus facilitating the investigation of subtle phenom-
ena not easily tested by means of a corpus search. For these rea-
sons, sentence judgment data have been crucial for the progress of
research in generative linguistics.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 771

Although the use of sentence judgments is still widely used


and accepted by the top journals in linguistics, Cowart (1997) and
Schütze (1996), among others, discuss a number of factors irrel-
evant to the linguistic phenomena under investigation that have
been shown to affect sentence judgments. These factors include facts
about the speakers themselves (e.g., researcher bias, dialect spoken),
the procedure used (e.g., ordering, repetition of sentences), and the
sentences chosen (e.g., lexical content, parsability, or pragmatic odd-
ity). Such factors call into question the validity of many studies in
generative syntax, particularly those studies that have investigated
subtle differences in acceptability and that have relied on the judg-
ments of only one or two speakers (often, the researcher or his/her
students).
Despite the many factors that can affect sentence judgments, both
Cowart (1997) and Schütze (1996) argue that such judgments can
provide reliable data even about very subtle syntactic phenomena if
extraneous factors are adequately controlled for. Both authors rec-
ommend using experimental methods similar to those routinely used
in cognitive psychological research. Among the key components of
such methods are: (1) an adequate number of participants to allow
a meaningful statistical analysis; (2) screening of participants for lin-
guistic background and other possible biases; (3) a factorial design
to distribute any extraneous factors as evenly as possible throughout
the set of test sentences; (4) procedures designed to avoid the effects
of repetition and ordering, and to avoid speakers guessing the exact
goal or expected outcome of the experiment.
During the course of our investigation of Cantonese coverbs, a
preliminary survey of three native speakers indicated that the inter-
actions between object extraction and aspect marking were too sub-
tle and variable to be understood by the traditional method of
informally interviewing a small number of speakers. Therefore, we
conducted a more carefully controlled sentence judgment experi-
ment, as described in the following sections.

4.2. Hypotheses
The experiment was designed to test four hypotheses based on our
analysis in Section 3 above.
(i) Coverbs are verbs (not prepositions) and should therefore
allow aspect marking. We predict that two-verb sentences
772 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

without extraction should be judged as highly acceptable


both with and without aspect marking on V1.
(ii) Aspect marking is normally indicated on the head verb of
a SVC (Déchaine 1993b, Li 1991). Therefore, aspect mark-
ing on V1 in a two-verb sentence may cause that verb to be
interpreted as the head of the SVC. Since the adjunct island
constraint would be inapplicable in such cases, we predict
that two-verb sentences with extraction should be judged as
significantly more acceptable when aspect is present on V1
than when no aspect is present on V1.
(iii) For simple one-verb clauses, extraction should be possi-
ble, since there is no constraint barring extraction from the
main verb. We predict that there should be no difference
in acceptability between simple clauses with extraction and
simple clauses without extraction.
(iv) An adjunct island constraint applies to phrases and clauses
functioning as adjuncts in the coverb construction. We pre-
dict that (a) two-verb clauses (coverb constructions) with-
out extraction should be significantly more acceptable than
two-verb clauses with extraction; (b) one-verb clauses with
extraction should be significantly more acceptable than
two-verb clauses with extraction (see hypothesis (iii) above).
If hypothesis (ii) is confirmed, these effects should show up
only in the no aspect group. Otherwise, these effects should
apply to both groups of subjects.

4.3. Methods
As noted above, three linguistic factors were manipulated in this
experiment: (1) presence vs. absence of aspect marking on V1;
(2) choice of verb in V1 position; and (3) sentence type. Aspect was
a between-groups factor with two levels (aspect/no aspect), while
verb and sentence type were within-group factors with six levels
(six different verbs) and four levels (four different sentence types),
respectively. Details of the participants, test materials, and proce-
dures are given below.

4.3.1. Participants
A total of 40 people participated in this study, with 20 assigned to
the ‘no aspect’ group and 20 to the ‘aspect’ group. Participants were
all adult native Cantonese speakers with normal speech and hearing
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 773

abilities, as established by a background questionnaire. They were


recruited using an email advertisement sent to several undergradu-
ate class email lists. All were undergraduate students from the Uni-
versity of Hong Kong with either no formal linguistics training or
an introductory course in linguistics, aged between 18 and 21. There
were 37 women and three men.10 Participants were informed of the
nature of the study, and each read and signed an informed consent
form prior to participation in the study. Each participant was paid
HK$50 (approximately US$7) for completing the study, which took
about 10–15 min.

4.3.2. Test materials


To test extraction, a number of filler-gap dependencies were avail-
able, such as topicalization, focusing, and relativization. Topicaliza-
tion and focusing were dispreferred since the test sentences were
to be presented out of context, making topicalization or focusing
possibly unmotivated and stylistically awkward. Relativization was
therefore chosen as a form of extraction which should be interpret-
able even out of context, as in (34).
(34) ??[[ngo pui-gwo i sik faan RC ] go go
I accompany-ASP eat rice that CL
haaki NP ] zau-zo.
person leave-ASP
Intended: ‘the person that I’ve had lunch with has left’
Six coverbs were selected as V1, representing a range from the most
verb-like (pui4 ‘accompany’) to the most preposition-like (tung4
‘with’) in terms of their meanings and traditional classifications:
(35) a. pui4 ‘accompany’ d. gan1 ‘follow, with’
b. doi6 ‘replace, act in place of’ e. hai2 ‘at’
c. wan2 ‘seek’ f. tung4 ‘with’
Although we expected no differences among verbs, we chose a range
of verbs so that any differences that might be present would be
likely to show up, thus providing a definitive test for our hypoth-
esis that (all) coverbs are verbs (not prepositions) and should allow
10 This proportion is representative of the proportion of women to men in the
Faculty of Arts at the University of Hong Kong, where about 90% of undergradu-
ates are women. We would, however, have no reason to expect any effect of gender
on sentence understanding in Cantonese.
774 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

aspect marking when used in the coverb construction (see hypothe-


sis (i) above).
For each coverb, a token set including four sentence types was
constructed, as shown in Table I below. Listeners in the no-aspect
group heard and rated sentences of types A-D, while listeners in
the aspect group rated sentences of types E-H, which differ from
the A-D sentences only with respect to aspect marking. For sen-
tences presented to the aspect group, the aspect marker gwo3 was
chosen as one which fitted all the contexts semantically, and had
been judged acceptable (in sentences without extraction) in the pilot
study. This morpheme marks ‘experiential’ aspect, denoting an event
which has happened at least once before (and has therefore been
‘experienced’). Thus, four sentence types (A-D or E-H) with each of
the six verbs were used for each group, making a total of 24 test
sentences per participant.
Following Cowart (1997), the experiment was designed to avoid
certain confounding factors. To avoid the participants focusing on
the experimental stimuli, a larger number of ‘distractor sentences’
(32) were mixed in with the 24 test sentences, making a total of
56 sentences per participant. These distractor sentences varied in
acceptability from highly acceptable to highly unacceptable along
a number of orthogonal syntactic dimensions. To avoid ordering
effects, the test and distractor sentences were presented in random
order. In addition, eight ‘practice sentences’ varying in acceptability
were presented before starting the experiment to familiarize partic-
ipants with the task. Participants were divided into aspect and no
aspect groups to avoid participants hearing the same sentence twice
(i.e., with and without aspect marking on V1), since multiple pre-
sentations of the same sentence have been shown in some studies to
result in familiarity effects (Schütze 1996: 135–140). See Appendix
A for a complete list of the test sentences.

4.3.3. Procedure
The sentences were recorded in a quiet room by a female native
speaker on a Sony Professional Walkman using an Aiwa micro-
phone. The test items were recorded in random order, to avoid
list intonation, and the recordings digitized. The recorded sentences
were presented to participants through Sennheiser HD 550 head-
phones on a Macintosh iBook laptop computer using an experiment
control program written in Hypercard. This program also collected
the responses automatically. Participants were given instructions in
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 775
TABLE I
Sentence types and examples

Sentence types Example token set for doi6 ‘replace’

A. Basic 1-verb Ngo doi go go tungsi


I replace that CL colleague
‘I’m taking the place of that colleague.’
B. Relativized Ngo doi go go tungsi beng-zo
1-verb I replace that CL colleague sick-ASP
‘The colleague whose place I’m taking is sick.’
C. Basic 2-verb Ngo doi go go tungsi gaau syu
I replace that CL colleague teach book
‘I’m taking the place of that colleague to
teach the class.’
D. Relativized Ngo doi gaau syu go go tungsi
2-verb I replace teach book that CL colleague
beng-zo
sick-ASP
‘The colleague whose place I’m taking to teach the
class is sick.’
E. Basic 1-verb Ngo doi-gwo go go tungsi
with aspect I replace-ASP that CL colleague
‘I’ve taken the place of that colleague.’
F. Relativized Ngo doi-gwo go go tungsi beng-zo
1-verb I replace-ASP that CL colleague sick-ASP
with aspect ‘The colleague whose place I’ve taken is sick.’
G. Basic 2-verb Ngo doi-gwo go go tungsi gaau syu
with aspect I replace-ASP that CL colleague teach book
‘I’ve taken the place of that colleague to teach the class.’
H. Relativized Ngo doi-gwo gaau syu go go
2-verb I replace-ASP teach book that CL
with aspect tungsi beng-zo
colleague sick-ASP
‘The colleague whose place I’ve taken to teach the
class is sick.’
776 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

Cantonese regarding the nature of the task and asked to enter a


judgment on a scale of 1 (‘good, completely acceptable’) to 4 (‘bad,
completely unacceptable’) after hearing each sentence.11 Judgments
were entered by clicking on the appropriate button on the screen,
and the ends of the scale were labeled as ‘good’ and ‘bad’.12 Partic-
ipants were allowed to take as much time as desired to submit their
judgment before the next sentence was presented, and the task was
completed within 10–15 min per participant.

4.4. Results
Results were analyzed using non-parametric statistics to test
the four hypotheses given in Section 4.2 above.13 Between-group
comparisons for hypotheses (i–ii) were calculated using the
Wilcoxon–Mann–Whitney Test (Siegel and Castellan 1988: 128–
137), while within-group comparisons for hypotheses (iii–iv) were
calculated using the Wilcoxon Signed Ranks Test (Siegel and Cas-
tellan 1988: 87–95).
Hypothesis (i): Confirmed. Both the aspect and the no aspect
groups judged two-verb sentences without extraction to be highly

11 An even-numbered ordinal scale was used so that subjects would be forced to


differentiate between ‘good’ (1 or 2) and ‘bad’ (3 or 4) sentences, while still allowing
subjects to indicate degree of goodness. See Appendix B for an English translation
of the instructions given to participants.
12 An anonymous reviewer notes that readers and subjects may find it counterin-
tuitive for a low-numbered rating to be more acceptable than a high-numbered rat-
ing. Although this may be true, it seems plausible that some people would find it
more intuitive to label an acceptable sentence with a low number, since we often
speak of the best person, team, or item as being number one, or first place. To
avoid this problem, the scale that the subjects used was clearly labeled with ‘good’
and ‘bad’ at each end, and there appear to be no cases in which subjects mistakenly
reversed the scale.
13 The experiment was originally designed with a standard analysis of variance
in mind. However, Bard et al. (1996: 40) point out that ordinal scales such as the
one used in this experiment are quite limited as to the kinds of statistics that can
be reliably applied to them, and they argue that ratio scales such as magnitude esti-
mation scales are to be preferred for studies using sentence judgments. Siegel and
Castellan (1988) further argue that parametric statistics of any kind are inappro-
priate for use with an ordinal scale. We have therefore limited our focus to those
hypotheses that can be reliably tested using non-parametric statistics: that is, we
have used tests that only measure rank ordering of variables and make no assump-
tions about the actual distance or the magnitude of difference between points on a
scale. We are grateful to an anonymous reviewer for alerting us to this problem.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 777
4.00
Bad

No aspect

3.50
With aspect

3.00
Rating

2.50

2.00

1.50
Good

1.00
Basic 1-Verb Basic 2-Verb Relativized 1- Relativized 2-
(A/E) (C/G) Verb (B/F) Verb (D/H)

Sentence Type
Figure 1. Acceptability ratings by sentence type with and without aspect.

acceptable. The mean rating for the no aspect group was 1.18, while
the mean rating for the aspect group was 1.14, as shown in the sec-
ond column of Figure 1 below. As expected, the difference between
the two groups was not significant (z = 0.23, p = 0.409). We also
note that although a range of verb-like and preposition-like coverbs
was used in the test sentences, aspect marking was highly acceptable
for all of the verbs tested.
Hypothesis (ii): Not supported. As predicted, two-verb sentences
with extraction were judged as slightly more acceptable when aspect
is present on V1 than when no aspect is present on V1: the mean
rating for the aspect group was 2.7, as compared with a mean rating
of 2.92 for the no aspect group (see the fourth column of Figure 1,
above). However, this difference was not significant (z = 0.73, p =
0.233).
Hypothesis (iii): Not supported. Simple clauses with extraction
were judged as less acceptable than simple clauses without extrac-
tion. The mean rating for simple clauses with extraction was 2.48,
while the mean rating for simple clauses without extraction was 1.95
(see Figure 2, first and third columns). This difference was signifi-
cant (z = 4.982, p < 0.001).
778 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

4.00
Bad

3.50

3.00
Rating

2.50

2.00

1.00
Good

1.00
Basic 1-Verb Basic 2-Verb Relativized 1-Verb Relativized2-Verb
Sentence Type
Figure 2. Acceptability ratings by sentence type.

Hypothesis (iv): Confirmed for parts (a) and (b). (a) Two-verb
clauses (coverb constructions) without extraction, with a mean rat-
ing of 1.16, were judged as more acceptable than two-verb clauses
with extraction, with a mean rating of 2.81 (see the second and
fourth columns of Figure 2). This difference was significant (z = 5.51,
p < 0.001). (b) One-verb clauses with extraction, with a mean rating
of 2.48, were judged to be more acceptable than two-verb clauses
with extraction, with a mean rating of 2.81 (see the third and fourth
columns of Figure 2). This difference was also significant (z = 3.017,
p = 0.0013). Note that because hypothesis (ii) was not supported, the
aspect and no aspect groups were combined for the purposes of these
tests.
Additional results: In examining the data, we noticed that the rat-
ings for two-verb sentences with extraction differed depending on
which verb was used in V1 position. Within the relativized two-
verb sentence type, the mean ratings for each verb were as follows:
pui3 ‘accompany’ = 2.53; doi6 ‘replace’ = 2.43; wan2 ‘seek’ = 2.80;
gan1 ‘follow’ = 2.20; hai2 ‘at’ = 3.38; and tung4 ‘with’ = 3.53 (see
Figure 3). Although all verbs in this sentence type showed marginal
or low acceptability, as expected, it appeared from the graph that the
sentences with hai2 ‘at’ and tung4 ‘with’ were less acceptable than
the sentences with the other four verbs. To test whether this differ-
ence might be significant, we conducted a Wilcoxon Signed Ranks
Test comparing the ratings for hai2 ‘at’ (i.e., the more acceptable of
Bad
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 779

4.00

3.50

3.00
Rating

2.50

2.00

1.50
Good

1.00
pui4 doi6 'replace' wan2 'seek' gan1 'follow' hai2 'at' tung4 'with'
'accompany'
Verb
Figure 3. Acceptability ratings by verb in relativized 2-verb sentences.

the two least acceptable verbs) with ratings for wan2 ‘seek’ (i.e., the
least acceptable of the remaining four verbs).14 The difference was
significant (z = 3.26, p < 0.001). Some possible implications of this
unexpected result are discussed in the following section.

4.5. Discussion
The main proposals from Section 3 are confirmed by the experimen-
tal results. First, all of the verbs we tested were highly acceptable
with aspect marking in the coverb construction (i.e., the two-verb
sentence type without extraction), confirming their verbal status
with respect to morphosyntax. Thus, the data appear to support
an analysis of coverbs as verbs rather than as prepositions. Sec-
ondly, as expected, two-verb sentences with extraction (i.e., coverbs
with relativized objects) were significantly less acceptable than the
same sentences without extraction. Similarly, two-verb sentences
with extraction were significantly less acceptable than one-verb (sim-
ple) sentences with extraction. Taken together, these results support
our hypothesis that an adjunct island constraint is responsible for
the incompatibility of coverb constructions with constructions such
as relative clauses, which involve extraction of the coverb’s object.
There were some results that were contrary to what we expected,
and these deserve some attention. First, the results of the experiment

14 Although post hoc tests would be more appropriate here if we were using
parametric statistics, there are no comparable post hoc tests using non-parametric
statistics.
780 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

failed to support our hypothesis that aspect-marking on V1 would


cause V1 to be parsed as the head of the coverb construction,
resulting in improved judgments of relativized two-verb sentences.
There was no significant difference between relativized two-verb
sentences with and without aspect, although there was a slight
difference in the predicted direction. There are at least two possible
explanations for this result:

(a) Aspect marking does not reliably indicate headedness in the


coverb construction, and therefore does not affect the appli-
cation of the proposed adjunct island constraint;
(b) Aspect marking indicates syntactic headedness but not
semantic headedness, and only semantic headedness is
relevant to the application of the adjunct island constraint.

The data are consistent with both possibilities. It is clear that V2 is


the semantic head of the coverb construction regardless of the posi-
tion of aspect marking. Since aspect marking takes semantic scope
over the entire SVC, and has the same semantic effect regardless of
which verb it attaches to, aspect marking has no apparent connection
with semantic headedness. However, it is still possible that aspect
marking indicates syntactic headedness independently of semantic
headedness, and that the adjunct island constraint is a constraint on
extraction from semantic modifiers (rather than syntactic adjuncts).
This would be similar to Culicover and Jackendoff’s (1997) formu-
lation of the Coordinate Structure Constraint as a constraint on
semantic well-formedness which applies to semantically coordinate
structures independently of whether these structures are also syntac-
tically coordinate. In the absence of evidence to distinguish these two
analyses, we will assume the simpler analysis in (a) above, where V2
is both the syntactic and the semantic head, and aspect marking is
independent of syntactic or semantic headedness.
A second unexpected result was that the verbs tung4 ‘with’ and
hai2 ‘at’ were found to be less acceptable in two-verb sentences with
extraction than the other four verbs. It might be tempting to sug-
gest that hai2 ‘at’ and tung4 ‘with’ are prepositions, and thus sub-
ject to a preposition-stranding constraint, while the other four items
are verbs. However, this would not explain why all six verbs allow
aspect marking in the two-verb sentences without extraction and
show degraded acceptability in the two-verb sentences with extrac-
tion. Another possibility would be to postulate some structural
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 781
TABLE II
Frequency of main verb vs. coverb (adjunct) uses of four verbs in the Cantonese
Radio Corpus

hai2 ‘at’ tung4 ‘with’ wan2 ‘seek’ gan1 ‘follow’

Main Verb 50 34% 2 1% 31 74% 4 67%


Coverb (V1) 59 41% 144 97% 11 26% 2 33%
Other 37 25% 3 2% 0 0% 0 0%
Total 146 100% 149 100% 42 100% 6 100%

difference between the sentences containing hai2 and tung4 and the
sentences containing the other verbs. However, we can find no inde-
pendent motivation for such a structural difference either. Instead,
we conjecture that the frequency of occurrence of each verb in par-
ticular syntactic contexts may have affected listeners’ judgments. A
search of the Cantonese Radio Corpus provides some preliminary
support for this hypothesis:15
As shown in Table II, hai2 ‘at’ and tung4 ‘with’ were used more
often as adjuncts (i.e., as V1 in a coverb construction), while wan2
‘seek’ and gan1 ‘follow’ were used more often as the main verb of
a simple clause.16 Based on these usage patterns, Cantonese speak-
ers may perceive hai2 and tung4 to be more prototypical examples
of adjuncts, whereas wan2 and gan1 are more prototypical exam-
ples of main verbs. Because the adjunct island constraint clearly
applies to hai2 and tung4, listeners may have been more confident
in their judgments of relativized two-verb sentences with hai2 and
tung4 as unacceptable. However, the corpus data are only suggestive,
and confirmation of this hypothesis awaits future research.
Finally, one-verb (simple) sentences with extraction were judged
to be significantly less acceptable than one-verb sentences with-
out extraction. Since the adjunct island constraint does not apply
to one-verb sentences, we did not expect to find any difference

15 The Cantonese Radio Corpus (Francis, Yiu, Matthews, and Chu, in prepara-
tion) is a database of colloquial spoken Cantonese from four radio talk programs
in Hong Kong, transcribed and tagged for syntactic analysis.
16 The corpus did not provide any useful data for the verbs pui4 ‘accompany’
and doi6 ‘replace’: there was only one example of pui4 ‘accompany’ in the cor-
pus, where it is used as the main verb of a simple clause, and no examples of doi6
‘replace’.
782 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

between these two sentence types. Although the difference between


one-verb and two-verb sentences with extraction can plausibly be
attributed to an adjunct island constraint, as originally hypothe-
sized, the difference between one-verb sentences with and without
extraction must be attributed to some other factor. One possibility
is that some of the verbs we used created a kind of garden path
effect in the one-verb sentences. Because we used the same verbs
in both the one-verb and two-verb sentences, we were only able to
choose those verbs that could be used as adjuncts in the coverb con-
struction. It is therefore possible that listeners initially analyzed the
verb of a simple one-verb clause as an adjunct. When listening to
(36), for example, subjects may have initially analyzed doi6 ‘replace’
as a coverb functioning as an adjunct.
(36) ngo doi go go tungsi beng-zo
I replace that CL colleague sick-ASP
‘The colleague that I replaced is sick.’
This suggestion receives some preliminary support from the fact
that for the verb wan2 ‘seek’, the ratings were identical for one-verb
sentences with and without extraction. Ratings were nearly identical
for gan1 ‘follow’ as well.17 As shown in Table II above, both wan2
and gan1 were used as main verbs more often than as coverbs in
the Cantonese Radio Corpus, suggesting that these verbs should be
less susceptible to misanalysis as adjuncts. In the absence of further
evidence, we will leave this as a hypothesis for further investigation.

5. F ORMULATING THE A DJUNCT I SLAND C ONSTRAINT

The extraction facts obtained in the sentence judgment experiment


support the analysis proposed in Section 3 above, whereby extrac-
tion is ruled out by an adjunct island constraint of some kind. In
this section, we will propose a specific formulation of the constraint.
We first argue that existing formulations of adjunct island con-
straints are inadequate to describe the relevant constraint in Can-
tonese. Secondly, we propose a simple, language-specific formulation
of the constraint. Thirdly, we suggest that Hawkins’ (1999) principle

17 The mean rating was 1.98 for one-verb sentences with wan2 both with and
without extraction. The mean rating was 1.43 for one-verb sentences with gan1
without extraction, and 1.48 for one-verb sentences with gan1 with extraction.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 783

of Avoid Competing Subcategorizers provides a plausible processing


motivation for the relevant constraint.

5.1. Existing Formulations: The CED, Subjacency, and Post-cyclic


Merger
Many languages prohibit extraction out of subordinate clauses func-
tioning as sentence adjuncts, as in the following English example:
(38) *Whoi did Mary cry after John hit i? (Huang 1998: 358)
While in some languages, including English and Mandarin, this
adjunct island constraint is limited to constructions involving “overt”
movement (Huang 1998), other languages, such as Lakhota, pro-
hibit extraction from subordinate clause adjuncts even in wh-in situ
constructions (Van Valin 1995). Cantonese is more permissive than
English or Mandarin in this respect: extraction out of a subordinate
clause functioning as an adjunct is acceptable in a relative clause or
topicalized sentence (where there is a gap or trace in the position of
the extracted phrase) as well as in a wh-question or with a resump-
tive pronoun (where there is no overt movement). That is, Cantonese
does not appear to display classic CED effects in cases of extraction
from a subordinate clause, as shown in (39a–c).
(39)a. go go jani , ngo gin jyun i zihau hou
that Cl person I see finish after very
satmong
disappointed
‘That person, I was disappointed after meeting.’
(Topicalization)
b. lei gin jyun i zihau hou satmong go
you see finish after very disappointed that
go jani . . .
Cl person
‘The person you were disappointed after meeting. . . ’
(Relativization)
c. lei gin jyun bingo zihau zeoi satmong?
You see finish who after most disappointed
‘Who were you most disappointed after meeting?’
(Wh-question)
784 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

Extraction from a coverb phrase shows a somewhat different pat-


tern: it is unacceptable in a relative clause or topicalized sentence,
but acceptable in a wh-question or with a resumptive pronoun, as
shown in (40a–c).
(40)a. lei tung-gwo bingo laamzai haang gungsi?
you with-ASP which boy walk shop
‘Which boy have you been out shopping with?’
(Wh-question)
b. *ngo tung-gwo i haang gungsi go go
I with-ASP walk shop that CL
laamzaii hou sing ge
boy very smart PRT
‘The boy I’ve been out shopping with is really smart.’
(Relativization)
c. ngo tung-gwo keoii haang gungsi go go
I with-ASP him walk shop that CL
laamzaii hou sing ge
boy very smart PRT
‘The boy I’ve been out shopping with is really smart.’
(Relativization with Resumptive Pronoun)
Thus, if the CED applies at all, it applies to overt movement oper-
ations involving coverb phrases functioning as adjuncts (as in 40)
but not to subordinate clauses functioning as adjuncts (as in 39).
It is therefore problematic to analyze Cantonese coverb phrases
as adjunct islands similar to subordinate clauses functioning as
adjuncts.
An additional problem is that existing formulations of the CED
and similar adjunct island conditions prove difficult to apply to the
Cantonese coverb construction. As discussed in Section 2, Huang’s
(1982) CED says that extraction out of a phrase or clause is only
possible if that phrase or clause is properly governed. Since adjuncts
are not the complement of any lexical head (and hence not properly
governed), extraction of a NP argument out of any kind of adjunct
violates the CED. This constraint would seem, then, to apply straight-
forwardly to the Cantonese coverb construction. However, we noted
in Section 2 that Huang’s (1998) formulation of the CED is prob-
lematic because the matrix VP of any sentence is not properly gov-
erned by a lexical head, and therefore extraction of the direct object
of the main verb (as in, What did she do?) should result in reduced
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 785

grammaticality, contrary to fact. Similarly, in the coverb construc-


tion, neither VP1 nor VP2 is properly governed, and yet extraction
from VP2 (the main verb phrase) is fully grammatical.
Chomsky (1986) subsumes the CED under the principle of Sub-
jacency, but essentially the same problem remains. The matrix VP
is not L-marked, and should therefore be a barrier to the extrac-
tion of its complement (1986: 29). He resolves this problem by
proposing a cyclic movement of the main verb’s complement first
to a position adjoined to VP before finally landing in Spec CP,
so that no more than one barrier is crossed for each movement.
In contrast, extraction from a subordinate clause functioning as
an adjunct must cross at least two barriers, resulting in the famil-
iar adjunct island effects exemplified in (38) above (Chomsky 1986:
31). It is far from clear, however, how a cyclic movement analysis
could account for the ungrammaticality of extraction from VP1 in
a Cantonese coverb construction. Assuming the constituent struc-
ture in (24) (repeated in (41) below), there is no reason why cyclic
movement should work differently in VP1 vs. VP2. For a cyclic
movement analysis to work, a different constituent structure would
have to be proposed – perhaps a structure in which VP1 is actu-
ally a subordinate clause with an empty subject. However, we can
find no independent motivation for complicating the analysis in this
way. Furthermore, such an analysis would predict that examples like
(39a–b) above should also be unacceptable, contrary to fact.
(41)

NP VP 2

keoi
s/he
VP 1 VP 2

V1 NP V2 NP

tung go go laamzai haang gungsi


with that CL boy walk shop
786 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

In a more recent formulation of adjunct island effects within the


Minimalist Program, Stepanov (2001a, b) proposes that adjunction is
a postcyclic merger operation. Stepanov’s proposal does not invoke
proper government to rule out extraction from adjuncts, or cyclic
movement to allow extraction from the matrix VP. Rather, adjuncts
are distinct from other types of phrases because they contain no
unchecked uninterpretable features, and for this reason they are only
merged with the rest of the sentence after feature-checking has taken
place. In an English sentence involving wh-movement, for example,
the Q feature (i.e., an abstract feature that takes scope over the matrix
clause and triggers wh-movement) cannot be satisfied when the wh-
phrase is embedded in an adjunct because adjuncts are only merged
with the rest of the sentence after feature-checking has taken place.
This would then imply that, in the coverb construction, VP2 con-
tains unchecked, uninterpretable features prior to merger with the
subject, whereas VP1 does not. Thus, VP1 can only be merged by
adjunction after all other merger operations have taken place, mak-
ing extraction from VP1 impossible. Given that Cantonese lacks any
case marking or subject–verb agreement, however, it is unclear which
uninterpretable features would be involved in the checking operation
or why VP1 should differ from VP2 with respect to feature check-
ing. Although it is possible to posit abstract features in VP2 with no
morphological realization, Stepanov’s (2001a) proposal, like that of
Chomsky (1986), seems unnecessarily complicated when applied to
the Cantonese coverb construction.

5.2. A Language-specific Constraint within a Processing-based Theory


of Filler-gap Dependencies
Hawkins (1999) argues that the kinds of island constraints typi-
cally proposed in the formal syntax literature, such as the CED
and Subjacency, fail to account for the full range of cross-linguistic
variation in permissible structures. For example, Swedish, Japanese,
and Akan permit violations of Subjacency (specifically, the Complex
Noun Phrase Constraint) which are not readily explained in terms of
parametric variation in what counts as a bounding node (1999: 262).
At the same time, Hawkins argues that such principles also fail to
account for certain well-attested typological patterns. For example,
they do not explain the typological patterns described by Keenan
and Comrie’s (1977) accessibility hierarchy for relative clause forma-
tion. In place of principles such as Subjacency, Hawkins proposes
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 787

that “languages have conventionalized permitted vs. non-permitted


gap and copy pronoun environments, defined in terms of different
filler types and node combinations. The grammatical possibilities
for each construction are constrained by the [implicational] hierar-
chies and explainable by processing complexity” (1999: 267). Hawkins
(1999) provides extensive evidence for this position, which we will not
attempt to summarize here.
Following Hawkins’ approach, the relevant island constraint for
the Cantonese coverb construction may be formulated more simply
in language-specific, construction-specific terms, as follows:
(42) Cantonese Coverb Extraction Constraint (CCEC): A filler
cannot be co-indexed with a gap that is contained within
a VP functioning as an adjunct to another VP within the
same clause.
The Cantonese Coverb Extraction Constraint (CCEC) is specific
enough to rule out extraction from VP1 of a coverb construction with-
out ruling out extraction from an adjunct clause, and without requir-
ing any complicated machinery, such as additional structure in VP1 to
create more barriers (Chomsky’s Subjacency account), or additional
abstract features in VP2 (Stepanov’s feature-checking account).
One might object that the constraint in (42) is not general
enough to have any kind of explanatory value. However, Hawkins
shows that language-specific conventions such as these, which he
argues are necessary to account for the full range of cross-linguistic
variation in the distribution of gaps, conform to universal process-
ing principles in predictable ways. One such principle is as follows:
(43) Avoid Competing Subcategorizers: The human processor
prefers to avoid garden paths that result from competing
subcategorizers within an FGD. (Hawkins 1999: 277)
The idea is that, when there is more than one possible subcatego-
rizer within an FGD (Filler-Gap Domain),18 people may have more
difficulty identifying the correct location of the gap.19

18 An FGD is defined as the domain that includes all of the nodes between the
subcategorizer and its gap, as well as additional nodes that function as the gap’s
“dependent or disambiguating arguments” (Hawkins 1999: 248).
19 Work by Lyn Frazier and colleagues provides empirical support for the
hypothesis that sentences with multiple potential subcategorizers are in fact more
difficult to process, see e.g. Frazier, Clifton and Randall (1983). Some speakers have
788 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

(44) a. [Which student] did you ask Mary about ?


b. [About which student] did you ask Mary ?
(Hawkins 1999: 277)
Hawkins notes that in (44a), there are two possible subcategorizers
for the NP which student – the verb ask and the preposition about.
But in the pied-piping structure in (44b), the verb ask is the only
possible subcategorizer for the PP gap corresponding to about which
student. The principle of Avoid Competing Subcategorizers predicts
that (44a) should be more difficult to process than (44b) due to the
potential ambiguity as to the position of the gap. Although sen-
tences like (44a) are permitted in English (attesting to the language-
specific nature of many extraction constraints), most languages
with prepositions have a preposition-stranding constraint which pre-
vents the formation of sentences like (44a). Hawkins suggests that
preposition-stranding constraints are motivated (but not demanded)
by this principle. By eliminating prepositions as a possible type of
subcategorizer, many grammars have developed a conventionalized
way of avoiding competing subcategorizers within an FGD so that
ambiguities as in (44a) do not arise.
Importantly, there appear to be no languages in which preposi-
tions are possible subcategorizers for a gap, but verbs are not. Why
should this be the case? Hawkins gives two main reasons. One rea-
son is that prepositional gaps are often more deeply embedded than
verbal gaps, violating a principle called Minimize FGDs (Filler-Gap
Domains). For example, in (44a), the gap is contained within the PP
complement of ask, which is more deeply embedded than, for exam-
ple, the gap in (45). This results in a greater number of nodes within
the FGD for (44a) as compared with (45).
(45) Which student did you ask ?
Another factor is that verbs are more frequently occurring than prep-
ositions. Although at least one verb is required in most clauses, prep-
ositional phrases are usually optional, and therefore less frequent.
Hawkins concludes: “Hence, the more complex and less frequently
occurring gap option is eliminated in most grammars” (1999: 277).
We have already argued that coverbs are not prepositions, and that
a preposition-stranding constraint is inadequate to account for the

Footnote 19 continued
the intuition that (44b) is less natural than (44a) because of a stylistic dispreference
for pied-piping in spoken English.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 789

extraction facts surrounding Cantonese coverb constructions. How-


ever, we propose that the constraint in (42), like preposition-stranding
constraints in many other languages, is motivated by the principle of
Avoid Competing Subcategorizers. In a SVC with multiple transitive
verbs, such as the coverb construction, there are potentially as many
subcategorizers for a gap as there are verbs. By eliminating adjunct VPs
as potential subcategorizers, the grammar of Cantonese thus avoids
what would otherwise be a potential ambiguity as to the location of
the gap site. Why should it be that adjuncts rather than main verbs are
eliminated? Like prepositions in other languages, coverbs functioning
as adjuncts are optional, and thus occur less frequently than main verbs.
For this reason, coverbs functioning as adjuncts are more dispensable
as subcategorizers than main verbs are.
It should be noted that the constraint in (42) is not maximally
efficient at eliminating competing subcategorizers. For example, con-
sider the following topicalized sentence:
(46) [ni go mantaiNP ]i ngo tung keoi touleon-gwo i
this CL problem I with 3sg discuss-ASP
‘This problem, I’ve discussed with him.’
The constraint in (42) requires the listener to determine that the first
VP tung keoi ‘with him’ is indeed an adjunct, and not the matrix VP,
thus ruling out V1 as a possible subcategorizer. In (46) this is a rela-
tively easy task, since the verb tung4 ‘with’ is almost always used as
an adjunct and very rarely as a main verb (see Section 4.4, Table II).
However, in the case of coverbs which are commonly used as main
verbs (e.g., wan2 ‘seek’), the listener will not necessarily rule out V1
as the main verb (and subcategorizer of the gap) until hearing the
remainder of the sentence. Despite this, the task of identifying VP1
as an adjunct is made easier by the fact that only a small number of
verbs can occur in the V1 position of a coverb construction (see (2)
in Section 1). Although some coverbs are commonly used as main
verbs, the reverse is not the case. That is, most verbs in Cantonese
are never used in V1 position of a coverb construction.
For the constraint in (42) to be maximally efficient, all of the pos-
sible coverbs would need to be used exclusively as adjuncts so that
the listener knows right away that they are not possible subcategoriz-
ers for a gap. Processing efficiency may then be a motivating factor
in grammaticalization, consistent with Lord’s (1993) general obser-
vation that main verbs often develop into more limited serial verbs,
before finally forming a separate class of prepositions. Cantonese
790 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

coverbs are intermediate in this transition, with some items having


grammaticalized more fully than others. For example, tung4 ‘with’ is
much more limited in its distribution than wan2 ‘seek’, and therefore
more readily identifiable as an adjunct, hence our finding that tung4
is the least acceptable item in two-verb relativization (Figure 3). If all
of the coverbs eventually become limited exclusively to the V1 posi-
tion of a coverb construction, it should become easier for listeners
to exclude coverbs as possible subcategorizers for a gap. And once
the distribution of these verbs is limited to V1 position, coverbs may
also lose the morphosyntactic properties of verbs, leading to the for-
mation of a class of true prepositions. At this point, the constraint
in (42) is likely to be reanalyzed as an ordinary preposition-stranding
constraint, so that the listener only needs to identify the word class
of an item to satisfy Avoid Competing Subcategorizers.
In sum, adjunct island constraints such as the CED, preposition-
stranding constraints, and the CCEC must be formulated differ-
ently to account for the facts of individual languages. Nevertheless,
these constraints share some common motivations. Both preposition-
stranding constraints and the CCEC can increase processing effi-
ciency by eliminating a class of items as possible subcategorizers for
a gap (Avoid Competing Subcategorizers). From this perspective, the
CCEC is more similar to a preposition-stranding constraint than it is
to the CED, even though it resembles the CED in its formulation as
a kind of adjunct island constraint. The CED (or whatever accounts
for CED effects) and the CCEC do appear to share some common
motivations, however. For example, Hawkins proposes that gap sites
within phrases functioning as adjuncts are more difficult to process
than gap sites within phrases functioning as subcategorized comple-
ments because “there is less conventionalized linguistic knowledge
to make clear how the categories in the FGD fit together” (Haw-
kins 2004: 213). This processing principle would apply to any kind
of adjunct, including adjunct clauses and coverb phrases. In addi-
tion, Goldberg (2006: 129–151) proposes that many phrases func-
tioning as adjuncts are resistant to extraction because they are back-
grounded in discourse and therefore pragmatically incompatible with
the discourse-prominent positions of extracted elements. She argues
that this pragmatic constraint helps motivate numerous island effects
involving backgrounded constituents.20 Both of these proposals pro-
vide general motivations for adjunct island constraints, and could
20 For a similar proposal, see also Van Valin (1995).
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 791

in fact play a role in motivating both the CED (in languages such
as English) and the CCEC. However, we maintain that the principle
of Avoid Competing Subcategorizers (Hawkins 1999) provides the
clearest motivation for the CCEC because it applies to adjunct VPs
such as coverb phrases but not to subordinate clauses functioning as
adjuncts (which do allow extraction in Cantonese).21 A more com-
plete account of extraction from different kinds of Cantonese adjunct
phrases awaits further research.

6. C ONCLUSION

In Section 3, we argued that the verbal properties of coverbs reflect


their syntactic categorization as verbs, and we presented evidence that
coverb phrases function as adjuncts to the main verb – a function
often performed by prepositional phrases in other languages. The
experimental data reported in Section 4 confirm our analysis of co-
verbs as verbs: coverbs were judged to be highly acceptable both with
and without aspect marking in two-verb sentences without extrac-
tion. It is therefore plausible that the relevant constraint on extrac-
tion is an adjunct island constraint, as argued by Law (1996) and Li
(1991) for Mandarin, rather than a preposition-stranding constraint.
The extraction facts obtained in the experiment further support this
analysis: listeners judged two-verb clauses with extraction (for which
the adjunct island constraint would be applicable) as significantly less
acceptable than either one-verb clauses with extraction or two-verb
clauses without extraction (for which the adjunct island constraint
would be inapplicable). Finally, in Section 5, we considered how best
to formulate the proposed adjunct island constraint. After consider-
ing and rejecting three existing adjunct island constraints, we pro-
posed a simple, language-specific formulation of the constraint to
account for the unique properties of the Cantonese coverb construc-
tion. We argue that this constraint is motivated by the processing
principle that Hawkins (1999) calls Avoid Competing Subcategoriz-
ers – one of the same principles that motivates preposition-stranding
constraints in many other languages. Thus, while object extraction
from coverb constructions is prohibited by a kind of adjunct island
constraint, the function of the constraint in processing efficiency is
similar to that of a preposition-stranding constraint.

21 In a sentence such as (39a) above, for example, the first verb is the only tran-
sitive verb and therefore the only potential subcategorizer for the gap.
APPENDIX A: TEST SENTENCES 792
Token set 1: pui4 ‘accompany’

ID # Sentence Mean Standard Standard


rating error deviation
1A Ngo pui go go haak 1.75 0.20 0.91
I accompany that CL client
‘I accompany that client’
1B Ngo pui go go haak zau-zo 2.10 0.24 1.07
I accompany that CL client leave-ASP
‘The client I accompanied has left’
1C Ngo pui go go haak sik faan 1.10 0.07 0.31
I accompany that CL client eat rice
‘I accompany that client to have a meal’
1D Ngo pui sik faan go go haak zau-zo 2.70 0.18 0.80
I accompany eat rice that CL client leave-ASP
‘The client I accompanied to have a meal has left’
1E Ngo pui-gwo go go haak 1.35 0.15 0.67
I accompany-ASP that CL client
‘I’ve accompanied that client before’
1F Ngo pui-gwo go go haak zau-zo 2.65 0.25 1.14
I accompany-ASP that CL client leave-ASP
‘The client I accompanied before has left’
1G Ngo pui-gwo go go haak sik faan 1.00 0 0
ELAINE J. FRANCIS AND STEPHEN MATTHEWS

I accompany-ASP that CL client eat rice


‘I’ve accompanied that client to have a meal before’
1H Ngo pui-gwo sik faan go go haak zau-zo 2.35 0.20 0.88
I accompany-ASP eat rice that CL client leave-ASP
‘The client I accompanied to have a meal before has left’
Token set 2: doi6 ‘replace’

ID # Sentence Mean Standard Standard


rating error deviation

2A Ngo doi go go tungsi


I replace that CL colleague
‘I’m taking the place of that colleague’ 1.90 0.19 0.85
2B Ngo doi go go tungsi beng-zo
I replace that CL colleague sick-ASP
‘The colleague whose place I’m taking is sick’ 2.60 0.24 1.10
2C Ngo doi go go tungsi gaau syu
I replace that CL colleague teach book
‘I’m teaching a class in place of that colleague’ 1.25 0.16 0.72
2D Ngo doi gaau syu go go tungsi beng-zo
I replace teach book that CL colleague sick-ASP
‘The colleague in whose place I’m teaching a class is sick’ 2.40 0.23 1.05
2E Ngo doi-gwo go go tungsi
I replace-ASP that CL colleague
‘I’ve taken the place of that colleague before’ 1.65 0.17 0.75
2F Ngo doi-gwo go go tungsi beng-zo
I replace-ASP that CL colleague sick-ASP
‘The colleague whose place I’ve taken before is sick’ 3.00 0.13 0.56
2G Ngo doi-gwo go go tungsi gaau syu
I replace-ASP that CL colleague teach book
‘I’ve taught a class in place of that colleague before’ 1.30 0.16 0.73
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE

2H Ngo doi-gwo gaau syu go go tungsi beng-zo


I replace-ASP teach book that CL colleague sick-ASP
‘The colleague in whose place I’ve taught a class before is sick’ 2.45 0.28 1.23
793
Token set 3: wan2 ‘seek’
794
ID # Sentence Mean Standard Standard
rating error deviation

3A Keoi wan go go pangjau


S/he seek that CL friend
‘She’s looking for that friend’ 1.50 0.14 0.61
3B Keoi wan go go pangjau m haidou
S/he seek that CL friend not here
‘The friend she’s looking for is not here’ 1.10 0.07 0.31
3C Keoi wan go go pangjau king-gai
S/he seek that CL friend chat
‘She’s looking for that friend to chat with’ 1.30 0.11 0.47
3D Keoi wan king-gai go go pangjau m haidou
S/he seek chat that CL friend not here
‘The friend she was looking for to chat with is not here’ 3.00 0.15 0.65
3E Keoi wan-gwo go go pangjau
S/he seek-ASP that CL friend
‘She was looking for that friend before’ 1.35 0.13 0.59
3F Keoi wan-gwo go go pangjau m haidou
S/he seek-ASP that CL friend not here
‘The friend she was looking for before is not here’ 1.85 0.17 0.75
ELAINE J. FRANCIS AND STEPHEN MATTHEWS

3G Keoi wan-gwo go go pangjau king-gai


S/he seek-ASP that CL friend chat
‘She was looking for that friend to chat with before’ 1.15 0.08 0.37
3H Keoi wan-gwo king-gai go go pangjau m haidou
S/he seek-ASP chat that CL friend not here
‘The friend she was looking for to chat before with is not here’ 2.60 0.21 0.94
Token set 4: gan1 ‘follow’

ID # Sentence Mean Standard Standard


rating error deviation

4A A-Man gan go go sinsaang


Ah Man follow that CL teacher
‘Ah Man is (studying) with that teacher’ 2.25 0.18 0.79
4B A-Man gan go go sinsaang sing Wong ge
Ah Man follow that CL teacher name Wong PRT
‘The teacher Ah Man is (studying) with is called Wong’ 1.85 0.17 0.75
4C A-Man gan go go sinsaang hok kam
Ah Man follow that CL teacher learn piano
‘Ah Man is learning the piano with that teacher’ 1.15 0.08 0.37
4D A-Man gan hok kam go go sinsaang sing Wong ge
Ah Man follow learn piano that CL teacher name Wong PRT
‘The teacher Ah Man is learning the piano with is called Wong’ 2.20 0.21 0.95
4E A-Man gan-gwo go go sinsaang
Ah Man follow-ASP that CL teacher
‘Ah Man has (studied) with that teacher’ 1.70 0.19 0.86
4F A-Man gan-gwo go go sinsaang sing Wong ge
Ah Man follow-ASP that CL teacher name Wong PRT
‘The teacher Ah Man has studied with is called Wong’ 2.10 0.20 0.91
4G A-Man gan-gwo go go sinsaang hok kam
Ah Man follow-ASP that CL teacher learn piano
‘Ah Man has learnt the piano with that teacher’ 1.10 0.07 0.31
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE

4H A-Man gan-gwo hok kam go go sinsaang sing Wong ge


Ah Man follow-ASP learn piano that CL teacher name Wong PRT
‘The teacher Ah Man has learnt the piano with is called Wong’ 2.20 0.22 1.01
795
Token set 5: hai2
796
ID # Sentence Mean Standard Standard
rating error deviation

5A A-Maa jigaa hai go gaan sezilau


Mum now at that CL office
‘Mum is in that office now’ 1.20 0.12 0.52
5B A-Maa jigaa hai go gaan sezilau hou daai gaan
Mum now at that CL office very big CL
‘The office Mum is in now is really big’ 2.65 0.22 0.99
5C A-Maa jigaa hai go gaan sezilau gin haak
Mum now at that CL office see client
‘Mum is meeting clients in that office now’ 1.10 0.07 0.31
5D A-Maa jigaa hai gin haak go gaan sezilau hou daai gaan
Mum now at see client that CL office very big CL
‘The office Mum is meeting clients in now is really big’ 3.35 0.15 0.67
5E A-Maa hai-gwo go gaan sezilau
Mum at-ASP that CL office
‘Mum has been in that office before’ 2.70 0.21 0.92
5F A-Maa hai-gwo go gaan sezilau hou daai gaan
Mum at-ASP that CL office very big CL
‘The office Mum has been in is really big’ 3.05 0.18 0.83
ELAINE J. FRANCIS AND STEPHEN MATTHEWS

5G A-Maa hai-gwo go gaan sezilau gin haak


Mum at-ASP that CL office see client
‘Mum has met clients in that office before’ 1.10 0.07 0.31
5H A-Maa hai-gwo gin haak go gaan sezilau hou daai gaan
Mum at-ASP see client that CL office very big CL
‘The office Mum has met clients in is really big’ 3.40 0.15 0.68
Token set 6: tung4

ID # Sentence Mean Standard Standard


rating error deviation

6A Gaze wui tung go go laamzai


Sister will with that CL boy
‘Sister will be with that boy’ 3.15 0.21 0.93
6B Gaze wui tung go go laamzai hou lengzai
Sister will with that CL boy very good-looking
‘The boy sister will be with is very good-looking’ 3.60 0.13 0.60
6C Gaze wui tung go go laamzai haang gungsi
Sister will with that CL boy walk shop
‘Sister is going shopping with that boy’ 1.15 0.08 0.37
6D Gaze wui tung haang gungsi go go laamzai hou lengzai
Sister will with walk shop that CL boy very good-looking
‘The boy sister is going shopping with is very good-looking’ 3.85 0.08 0.37
6E Gaze tung-gwo go go laamzai
Sister with-ASP that CL boy
‘Sister has been with that boy’ 2.85 0.18 0.81
6F Gaze tung-gwo go go laamzai hou lengzai
Sister with-ASP that CL boy very good-looking
‘The boy sister has been with is very good-looking’ 3.20 0.19 0.83
6G Gaze tung-gwo go go laamzai haang gungsi
Sister with-ASP that CL boy walk shop
‘Sister has been shopping with that boy’ 1.20 0.09 0.41
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE

6H Gaze tung-gwo haang gungsi go go laamzai hou lengzai


Sister with-ASP walk shop that CL boy very good-looking
‘The boy sister has been shopping with is very good-looking’ 3.20 0.19 0.83
797
798 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

A PPENDIX B: I NSTRUCTIONS FOR PARTICIPANTS22

This is a survey about your intuitions regarding the acceptability


of certain sentences in Cantonese. There are no right or wrong
answers; we are only interested in your own opinions.
In a moment, you will hear a series of 8 Cantonese sentences, fol-
lowed by another series of 56 sentences. You will be asked to rate
each sentence on a scale of 1–4, where 1 means ‘good’ (completely
acceptable, an ordinary way of speaking), and 4 means ‘bad’ (com-
pletely unacceptable, people would never say it that way). Ratings of
2 or 3 indicate intermediate degrees of goodness.
After hearing each sentence, indicate your rating by clicking on
the appropriate number on the screen. When you are finished with
the study, please notify the laboratory assistant.

R EFERENCES

Andrews, Avery D. and Christopher D. Manning. 1999. Complex Predicates and


Information Spreading in LFG, CSLI, Stanford.
Baker, Mark C. 2003. Lexical Categories: Verbs, Nouns, and Adjectives, Cambridge
University Press, Cambridge.
Baltin, Mark. 1978. Constraints on Construal, Ph.D. dissertation, MIT, Cambridge,
MA.
Baltin, Mark R. and Paul M. Postal. 1996. ‘More on Reanalyses Hypotheses’, Lin-
guistic Inquiry 27, 127–145.
Bard, Ellen G., Dan Robertson, and Antonella Sorace. 1996. ‘Magnitude Estima-
tion of Linguistic Acceptability’, Language 72, 32–68.
Carstens, Vicki. 2002. ‘Antisymmetry and Word Order in Serial Constructions’,
Language 78, 3–50.
Chan, Alice Yin Wa. 2000. ‘Chinese Serial Verb Constructions with Experiential
Aspect Marker-Syntactic Representations and Semantic Interpretations’, Com-
munications of COLIPS 10, 69–97.
Chomsky, Noam. 1986. Barriers, MIT Press, Cambridge, MA.
Collins, Chris. 1997. ‘Argument Sharing in Serial Verb Constructions’, Linguistic
Inquiry 28, 461–497.
Cowart, Wayne. 1997. Experimental Syntax, Sage Publications, Thousand Oaks,
CA.
Croft, William. 1991. Syntactic Categories and Grammatical Relations, University of
Chicago Press, Chicago.

22 This is an English version of the instructions, which were actually read in


Cantonese by the Cantonese speaking RA who ran the experiment.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 799

Culicover, Peter W. and Ray S. Jackendoff. 1997. ‘Semantic Subordination Despite


Syntactic Coordination’, Linguistic Inquiry 28, 195–218.
Culicover, Peter W. and Ray S. Jackendoff. 2005. Simpler Syntax, Oxford University
Press, Oxford.
Dai, John Xiang-ling. 1990. ‘Syntactic Constructions in Serial Verb Expressions in
Chinese’, in B. Joseph and A. Zwicky (eds.), When Verbs Collide: Papers from the
1990 Ohio State Mini-Conference on Serial Verbs, Department of Linguistics, The
Ohio State University, Columbus, pp. 316–339.
Déchaine, Rose-Marie. 1993a. ‘Serial Verb Constructions’, in J. Jacobs et al.
(eds.), Syntax: An International Handbook of Contemporary Research, Walter de
Gruyter, Berlin, pp. 799–825.
Déchaine, Rose-Marie. 1993b. Predicates Across Categories, Ph.D. dissertation,
University of Massachusetts at Amherst.
Finney, Malcolm A. 1992. ‘The Theta Criterion and Argument-Sharing in Serial
Verb Constructions’, Cahiers Linguistiques d’Ottawa 20, 26–43.
Francis, Elaine J. 2005. ‘Syntactic Mimicry as Evidence for Prototypes in Gram-
mar’, in S. S. Mufwene, E. J. Francis, and R. S. Wheeler (eds.), Polymorphous
Linguistics: Jim McCawley’s Legacy, MIT Press, Cambridge, MA, pp. 161–181.
Francis, Elaine J. and Stephen Matthews. 2005. ‘A Multi-Dimensional Approach to
the Category ‘Verb’ in Cantonese’, Journal of Linguistics 41, 269–305.
Frazier, Lyn, Charles Clifton and Janet Randall. 1983. ‘Filling Gaps: Decision Prin-
ciples and Structure in Sentence Comprehension,’ Cognition 13, 187–222.
Goldberg, Adele E. 2006. Constructions at Work: The Nature of Generalization in
Language, Oxford University Press, Oxford.
Hagemeijer, Tjerk. 2001. ‘Underspecification in Serial Verb Constructions’, in
N. Corver and H. van Riemsdijk (eds.), Semi-lexical Categories: The Function of
Content Words and the Content of Function Words, Mouton de Gruyter, Berlin,
pp. 415–451.
Hawkins, John A. 1999. ‘Processing Complexity and Filler-Gap Dependencies
Across Grammars’, Language 75, 244–285.
Hawkins, John A. 2004. Efficiency and Complexity in Grammars, Oxford University
Press, Oxford.
Hopper, Paul J. and Sandra A. Thompson. 1984. ‘The Discourse Basis for Lexical
Categories in Universal Grammar’, Language 60, 703–752.
Hornstein, Norbert and Amy Weinberg. 1981. ‘Case Theory and Preposition
Stranding’, Linguistic Inquiry 12, 55–91.
Huang, C. T. James. 1982. Logical Relations in Chinese and the Theory of Grammar,
Ph.D. dissertation, MIT, Cambridge, MA.
Huang, C. T. 1998. Logical Relations in Chinese and the Theory of Grammar,
Garland Publishing, New York.
Jones, Charles. 1987. ‘P for Proper Governor’, West Coast Conference on Formal
Linguistics 115–130.
Keenan, Edward and Bernard Comrie. 1977. ‘Noun Phrase Accessibility and Uni-
versal Grammar’, Linguistic Inquiry 8, 63–99.
Koopman, Hilda. 1984. The Syntax of Verbs. Dordrecht, Foris.
Kwok, Fan, Thomas Lee, Caesar Lun, K. K. Luke, Peter Tung, and K. H. Cheung.
1997. Guide to LSHK Cantonese Romanization of Chinese Characters, Linguistic
Society of Hong Kong, Hong Kong.
800 ELAINE J. FRANCIS AND STEPHEN MATTHEWS

Law, Paul. 1996. ‘A Note on the Serial Verb Construction in Chinese’, Cahiers de
Linguistique – Asie Orientale 25, 199–233.
Lawal, Nike S. 1993. ‘Serial Verbs in Yoruba as Adjunct Phrases’, Lingua 91,
185–200.
Levine, Robert D. 1984. ‘Against Reanalysis Rules,’ Linguistic Analysis 14, 3–30.
Li, Charles and Sandra A. Thompson. 1981. Mandarin Chinese: A Functional Ref-
erence Grammar, University of California Press, Berkeley.
Li, Yafei. 1991. ‘On Deriving Serial Verb Constructions’, in C. Lefebvre (ed.), Ser-
ial Verbs: Grammatical, Comparative, and Cognitive Approaches, John Benjamins,
Amsterdam, pp. 103–136.
Li, Yafei. 1993. ‘Structural Head and Aspectuality’, Language 69, 480–504.
Li, Yen-hui Audrey. 1990. Order and Constituency in Mandarin Chinese, Kluwer,
Dordrecht.
Lord, Carol. 1993. Historical Change in Serial Verb Constructions, John Benjamins,
Amsterdam.
Matthews, Stephen. 2006. ‘On serial verbs in Cantonese’, in A. Y. Aikhenvald
and R. M. W. Dixon (eds.), Serial Verbs: a Cross-linguistic Typology, Oxford
University Press, Oxford, pp. 69–87.
Matthews, Stephen and Virginia Yip. 1994. Cantonese: A Comprehensive Grammar,
Routledge, London.
McCawley, James D. 1992. ‘Justifying Part-of-speech Assignments in Mandarin
Chinese’, Journal of Chinese Linguistics 20, 213–245.
Newmeyer, Frederick J. 1998a. ‘Preposition Stranding: Parametric Variation and
Pragmatics,’ Languages and Linguistics 1, 1–24.
Newmeyer, Frederick J. 1998b. Language Form and Language Function, MIT Press,
Cambridge, MA.
Rhys, Catrin Sian. 2000. ‘Chinese Coverbs and Argument Structure’, Linguistica
Atlantica 22, 69–87.
Ross, Claudia. 1991. ‘Coverbs and Category Distinctions in Mandarin Chinese’,
Journal of Chinese Linguistics 19, 79–114.
Ross, John Robert. 1967. Constraints on Variables in Syntax, Ph.D. dissertation,
MIT, Cambridge, MA.
Schütze, Carson. 1996. The Empirical Base of Linguistics: Grammaticality Judg-
ments and Linguistic Methodology, University of Chicago Press, Chicago.
Siegel, Sidney and N. John Castellan. 1988. Nonparametric Statistics for the Behav-
ioral Sciences, 2nd edn, McGraw Hill, Boston.
Stepanov, Arthur. 2001a. ‘The End of CED?’ in K. Megerdoomian and
L. A. Bar-el (eds.), Proceedings of the West Coast Conference on Formal
Linguistics, Vol. 20, Cascadilla Press, Somerville, MA, pp. 524–537.
Stepanov, Arthur. 2001b. ‘Late Adjunction and Minimalist Phrase Structure’, Syn-
tax 4, 94–125.
Van Valin, Robert D. 1995. ‘Toward a Functionalist Account of So-called Extrac-
tion Constraints’, in B. Devriendt, L. Goossens, and J. van der Auwera (eds.),
Complex Structures: A Functionalist Perspective, Mouton de Gruyter, Berlin,
pp. 29–60.
Veenstra, Tonjes. 2000. ‘Verb Serialization and Object Position’, Linguistics 38,
867–888.
CATEGORIALITY AND OBJECT EXTRACTION IN CANTONESE 801

Yuasa, Etsuyo and Jerrold M. Sadock. 2002. ‘Pseudo-Subordination: A Mismatch


Between Syntax and Semantics’, Journal of Linguistics 38, 87–111.
Zhang, Shi. 1990. ‘Correlations Between the Double Object Construction and Prep-
osition Stranding’, Linguistic Inquiry 21, 312–316.
Zwicky, Arnold. 1985. ‘Heads’, Journal of Linguistics 21, 1–30.

Received 29 July 2002


Revised 6 December 2005

Elaine J. Francis
Department of English
Purdue University
500 Oval Drive
West Lafayette, IN 47907
USA
<ejfranci@purdue.edu>

Stephen Matthews
Department of Linguistics
University of Hong Kong
Pokfulam Road, Hong Kong
China
<matthews@hkucc.hku.hk>

View publication stats

You might also like